Arpit, Devansh, Jastrzębski, Stanisław, Ballas, Nicolas, Krueger, David, Bengio, Emmanuel, Kanwal, Maxinder S., Maharaj, Tegan, Fischer, Asja, Courville, Aaron, Bengio, Yoshua, Lacoste-Julien, Simon

We examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While deep networks are capable of memorizing noise data, our results suggest that they tend to prioritize learning simple patterns first. In our experiments, we expose qualitative differences in gradient-based optimization of deep neural networks (DNNs) on noise vs. real data. We also demonstrate that for appropriately tuned explicit regularization (e.g., dropout) we can degrade DNN training performance on noise datasets without compromising generalization on real data. Our analysis suggests that the notions of effective capacity which are dataset independent are unlikely to explain the generalization performance of deep networks when trained with gradient based methods because training data itself plays an important role in determining the degree of memorization.

Yun, Chulhee, Sra, Suvrit, Jadbabaie, Ali

We study universal finite sample expressivity of neural networks, defined as the capability to perfectly memorize arbitrary datasets. For scalar outputs, existing results require a hidden layer as wide as $N$ to memorize $N$ data points. In contrast, we prove that a 3-layer (2-hidden-layer) ReLU network with $4 \sqrt {N}$ hidden nodes can perfectly fit any arbitrary dataset. For $K$-class classification, we prove that a 4-layer ReLU network with $4 \sqrt{N} + 4K$ hidden neurons can memorize arbitrary datasets. For example, a 4-layer ReLU network with only 8,000 hidden nodes can memorize datasets with $N$ = 1M and $K$ = 1k (e.g., ImageNet). Our results show that even small networks already have tremendous overfitting capability, admitting zero empirical risk for any dataset. We also extend our results to deeper and narrower networks, and prove converse results showing necessity of $\Omega(N)$ parameters for shallow networks.

Deep neural networks (DNNs) can easily fit a random labeling of the training data with zero training error. What is the difference between DNNs trained with random labels and the ones trained with true labels? Our paper answers this question with two contributions. First, we study the memorization properties of DNNs. Our empirical experiments shed light on how DNNs prioritize the learning of simple input patterns. In the second part, we propose to measure the similarity between what different DNNs have learned and memorized. With the proposed approach, we analyze and compare DNNs trained on data with true labels and random labels. The analysis shows that DNNs have \textit{One way to Learn} and \textit{N ways to Memorize}. We also use gradient information to gain an understanding of the analysis results.

Recurrent neural networks (RNNs) have been drawing much attention with great success in many applications like speech recognition and neural machine translation. Long short-term memory (LSTM) is one of the most popular RNN units in deep learning applications. LSTM transforms the input and the previous hidden states to the next states with the affine transformation, multiplication operations and a nonlinear activation function, which makes a good data representation for a given task. The affine transformation includes rotation and reflection, which change the semantic or syntactic information of dimensions in the hidden states. However, considering that a model interprets the output sequence of LSTM over the whole input sequence, the dimensions of the states need to keep the same type of semantic or syntactic information regardless of the location in the sequence. In this paper, we propose a simple variant of the LSTM unit, persistent recurrent unit (PRU), where each dimension of hidden states keeps persistent information across time, so that the space keeps the same meaning over the whole sequence. In addition, to improve the nonlinear transformation power, we add a feedforward layer in the PRU structure. In the experiment, we evaluate our proposed methods with three different tasks, and the results confirm that our methods have better performance than the conventional LSTM.

The roles played by learning and memorization represent an important topic in deep learning research. Recent work on this subject has shown that the optimization behavior of DNNs trained on shuffled labels is qualitatively different from DNNs trained with real labels. Here, we propose a novel permutation approach that can differentiate memorization from learning in deep neural networks (DNNs) trained as usual (i.e., using the real labels to guide the learning, rather than shuffled labels). The evaluation of weather the DNN has learned and/or memorized, happens in a separate step where we compare the predictive performance of a shallow classifier trained with the features learned by the DNN, against multiple instances of the same classifier, trained on the same input, but using shuffled labels as outputs. By evaluating these shallow classifiers in validation sets that share structure with the training set, we are able to tell apart learning from memorization. Application of our permutation approach to multi-layer perceptrons and convolutional neural networks trained on image data corroborated many findings from other groups. Most importantly, our illustrations also uncovered interesting dynamic patterns about how DNNs memorize over increasing numbers of training epochs, and support the surprising result that DNNs are still able to learn, rather than only memorize, when trained with pure Gaussian noise as input.